Question/Issue:
I have seemingly successfully trained an audio classification model. The Live Classification seems to return 100% for my 3 classes which is great and using the Live Classification from my test sample set seems to work great. I build and download, add the library to arduino ide, copy the sample code from static buffer, i get the ‘raw’ data from the test sample i ran in Live Classification…and the results are different.
Project ID:
edwards09-project-1
Context/Use case:
Steps Taken:
- Classify audio
- test with Live Classification
- get RAW data from Live Classification sample
- Build and download to Arduino IDE
- Set the features[] to the downloaded RAW data
- Upload to esp32s3
Expected Outcome:
The Predictions should be the same
Actual Outcome:
Different Results
Reproducibility:
- [x ] Always
- [ ] Sometimes
- [ ] Rarely
Environment:
Using an esp32s3 N16R8.
Have INMP441 attached.
Did NOT record the audio with this mic because i can’t seem to get it connected to Edge Impulse (however, this shouldn’t matter for this example because i’m using static buffer).
Below is a screenshot of the Live Classification properly identifying that the audio is in the ‘pump’ class.
I can’t seem to paste the entirety of the RAW data…but this is the first little bit.
static const float features[] = {
-120, -176, -163, -191, -191, -206, -196, -209, -198, -206, -198, -201, -190, -179, -163, -151, -130, -110, -95, -96, -92, -96, -96, -104, -112, -116, -118, -119, -112,.........................................
};
Here are the results…
Timing: DSP 90 ms, inference 4 ms, anomaly 0 ms
Predictions:
background_noise: 0.23828
pump: 0.00000
silence: 0.76172